Social Isolation in America: Changes in Core Discussion Networks over Two Decades Miller McPherson University of Arizona and Duke University Lynn Smith-Lovin Duke University Matthew E. Brashears University of Arizona Have the core discussion networks of Americans changed in the past two decades? In 1985, the General Social Survey (GSS) collected the first nationally representative data on the confidants with whom Americans discuss important matters. In the 2004 GSS the authors replicated those questions to assess social change in core network structures. Discussion networks are smaller in 2004 than in 1985. The number of people saying there is no one with whom they discuss important matters nearly tripled. The mean network size decreases by about a third (one confidant), from 2.94 in 1985 to 2.08 in 2004. The modal respondent now reports having no confidant; the modal respondent in 1985 had three confidants. Both kin and non-kin confidants were lost in the past two decades, but the greater decrease of non-kin ties leads to more confidant networks centered on spouses and parents, with fewer contacts through voluntary associations and neighborhoods. Most people have densely interconnected confidants similar to them. Some changes reflect the changing demographics of the U.S. population. Educational heterogeneity of social ties has decreased, racial heterogeneity has increased. The data may overestimate the number of social isolates, but these shrinking networks reflect an important social change in America There are some things that we discuss only with people who are very close to us. These important topics may vary with the situation or the person — we may ask for help, probe for information, or just use the person as a sound- ing board for important decisions — but these are the people who make up our core network of confidants. How have these discussion networks of close confidants changed over the past two decades? We address that question here with data from a high-quality national probability survey that collected parallel data in 1985 and 2004. We find a remarkable drop in the size of core discussion networks, with a shift away from ties formed in neighborhood and com- munity contexts and toward conversations with close kin (especially spouses). Many more peo- ple talk to no one about matters they consider Please address correspondence to Miller Networks in Columbus, Ohio, and at the Social 354 AMERICAN SOCIOLOGICAL REVIEW important to them in 2004 than was the case two decades ago. Why is this question (and its disturbing answer) significant? Social scientists know that contacts with other people are important in both instrumental and socio-emotional domains (Fischer 1982; Lin 2001). The closer and stronger our tie with someone, the broader the scope of their support for us (Wellman and Wortley 1990) and the greater the likelihood that they will provide major help in a crisis (Hurlbert, Haines, and Beggs 2000). These are important people in our lives. They influence us directly through their interactions with us and indirect- ly by shaping the kinds of people we become (Smith-Lovin and McPherson 1993). Much of what we know about these core con- fidants comes from surveys that measure ego- centered networks — relationships from the point of view of a single person. These data describe the interpersonal environment of an individual. They allow us to measure the degree to which that person is directly connected to different parts of a social system and integrated into it at the individual level. Building on earlier network surveys (e.g., Fischer 1982; Verbrugge 1977; Wellman 1979), the General Social Survey (GSS) measured the national U.S. social system of ego-networks for the first time in 1985 (Burt 1984; Marsden 1987). Since then, our description of the core interpersonal environment for Americans has been frozen in the mid-1980s. Of course, one expects major social indicators to change slow- ly, if at all. There is evidence, however, that the structure of social relationships in the United States has shifted in recent decades. Putnam (1995; 2000), in particular, has heightened inter- est in networks by emphasizing links among interpersonal ties, voluntary association mem- bership, community well-being, and civic par- ticipation. He follows a rich tradition, dating back to de Tocqueville, in arguing that Ameri- cans' ties to other members of their communi- and Moody 2004) to crime (Sampson and Laub 1990). To assess social change in American discus- sion networks, we replicated the 1985 network questions in the 2004 GSS, using the same question wording and highly similar data col- lection procedures. In this article, we first out- line what we know about the characteristics of the GSS questions — what kinds of networks they tap, with what reliability and validity, and what kinds of issues they leave unanswered. We then compare the basic characteristics of these core discussion networks at the two time points, 1985 and 2004. ' Given that the differ- ences, especially in network size, are very large, we consider methodological factors that might be biasing our results, and we provide data on these issues when possible. Finally, we decom- pose the differences in major network charac- teristics into meaningful methodological and substantive sources. We conclude with a dis- cussion of the potential importance of our find- ings. CORE DISCUSSION NETWORKS: WHAT KINDS OF TIES ARE WE MEASURING? The Questions When researchers study interpersonal environ- ments, a key issue is what type of relationship they want to measure. The ideal, of course, would be to assess several types of relationship (e.g., friend, coworker, advisor) and then to use those multi-layered data to find common pat- terns (see Fischer 1982 for an excellent exam- ple). Given the time constraints of a national face-to-face survey, the 1985 GSS instead 1 The GSS asked the same question in 1987 as part of a module on political participation. In 1987, how- ever, the survey did not ask sociodemographic char- acteristics and interconnections among network alters. The only network tie characteristics assessed were SOCIAL ISOLATION IN AMERICA 355 focused on a relation that was general, cogni- tively definable, and significant: it asked peo- ple with whom they discussed personally important topics. In his earlier study of California communi- ties, Fischer (1982) used a similar question about discussing personal matters. He found that this relationship elicited relatively strong personal ties with a good representation of both kin and non-kin. These close relationships have theoretical importance because they are cen- tral in social influence and normative pressures (Burt 1984:127), and have strong conceptual connections to earlier survey measures of best friends and other close socio-emotional ties. Different ways of asking about important, close interpersonal relationships (often called strong ties) tend to be convergent. 2 Many ways of ask- ing such questions get the same close ties. These close ties are only a small subset of a person's complete interpersonal environment, which also includes a much larger array of weak ties, which are more distant connections to peo- ple. Weak ties may occur in just one institutional context or may connect us to people who are less like us in many ways (demographically, politi- cally, or culturally; see Granovetter 1973; McPherson and Smith-Lovin 1981). Estimates of the larger network of weak ties range between 150 (Hill and Dunbar 2003) to more than a thousand (see review in Marsden 2005). In 2004, we replicated a substantial subset of the network questions. Specifically, the 1985 and 2004 GSS asked the following questions: From time to time, most people discuss important matters with other people. Looking back over the last six months — who are the people with whom you discussed matters important to you? Just tell me their first names or initials. IF LESS THAN 5 NAMES MENTIONED, PROBE: Anyone else? Please think about the relations between the people you just mentioned. Some of them may be total strangers in the sense that they wouldn't rec- ognize each other if they bumped into each other on the street. Others may be especially close, as Are they especially close? PROBE: As close or closer to each other as they are to you? The survey then asked about demographic characteristics of the discussion partner: whether the partner was male or female, his or her race, his or her education and age, and some aspects of the respondents' relationship with the dis- cussion partner. Then, the interviewer asked more about the character of the relationship: Here is a list of some of the ways in which people are connected to each other. Some people can be connected to you in more than one way. For exam- ple, a man could be your brother and he may belong to your church and be your lawyer. When I read you a name, please tell me all of the ways that person is connected to you. How is (NAME) connected to you? PROBE: What other ways? (The options were presented on a card: Spouse, Parent, Sibling, Child, Other family, Co-worker, Member of group, Neighbor, Friend, Advisor, Other). What These Questions Measure (and Miss) People's reports of their connections with other people are not perfect reflections of their actu- al interactions (Bernard, Killworth, and Sailer 1 982). On the other hand, people are quite good at remembering long-term or typical patterns of interaction with other people (Freeman, Romney, and Freeman 1987). Therefore, answers that respondents give to network ques- tions on surveys often represent their typical interpersonal environment rather than whatev- er researchers specifically asked them. As one might expect, respondents report frequently contacted, close, core network ties with those whom they have many types of relationships more reliably than they do more distant, simple relations (Kogovsek and Ferligoj 2004). These close ties are also more accessible in memory and tend to be listed first in a survey response (Brewer 1995; Burt 1986; Verbrugge 1977). Snrin-pmnrinnal tips tpnri tn he nampri mnre 356 AMERICAN SOCIOLOGICAL REVIEW complete network list generated with the help of extensive probes. The people most likely to be mentioned in the GSS question are strong, close ties who are more connected to others in the network (because one name helps the respondent to recall others). Ruan (1998) exam- ined the overlap of names generated by the GSS question and other network questions based on exchange relations in China. She found that the GSS discussion question accounted for an important part of a Chinese personal network. The people with whom the Chinese respon- dents discussed important matters were also likely to spend leisure time with the respon- dents and to be their confidants for personal matters. The respondents expected them to offer substantial help or to possess important social resources. Similarly, Burt (1997), in a study of managers, found that the GSS question elicit- ed high overlap with people whom the managers socialized with informally and considered their most valued contacts, and who they would want to ask for advice if they were considering a job change. These findings reinforce our sense that the important-matters question elicits the core, frequently accessed interpersonal environments that people use for sociality and advice, and for socio-emotional and instrumental support. While clarifying what the GSS question measures, we should also be clear about what it does not measure. Most obviously, it does not measure what people talk about in their relationships. Several studies have asked this interesting question to help fill in the content behind these discussion networks (Bailey and Marsden 1999; Bearman and Parigi 2004; Straits 2000). The studies agree that important matters vary dramatically from respondent to respondent, encompassing relevant personal matters (intimate relationships, finances, health, hobbies, and work problems), as well as more general topics such as political issues. They also agree that there are significant respon- dent-alter interactions in what types of topics arp rnrwirlprpd imnnrtnnt (Rparman anr\ Pnriai about a specific instance of discussion of a par- ticular important matter. Reinforcing this view, Bearman and Parigi (2004) found that some people cited apparently mundane matters like getting a hair cut when asked the topic of their latest discussion about important matters. Luckily, Bailey and Marsden (1999) also found that measures of key network characteristics (e.g., density, range, heterogeneity) tended not to vary across different interpretations of the question. In summary, the GSS network question about those with whom one discusses important mat- ters elicits a close set of confidants who are probably routinely contacted for talk about both mundane and serious life issues, whatever those might be for a given respondent. They represent an important interpersonal environment for the transmission of information, influence, and sup- port. We would be unwise to interpret the answers to this question too literally (e.g., assuming that a specific conversation about some publicly weighty matter had occurred in the past six months), but these answers do give us a window into an important set of close, rou- tinely contacted people who make up our respondents' immediate social circle. DATA AND ANALYSES The GSS is a face-to-face survey of the non- institutionalized United States adult popula- tion. 3 The 1985 and 2004 surveys used the same questions to generate the names of confidants and identical procedures to probe for addition- al discussion partners. Therefore, the survey responses represent a very close replication of the same questions and procedures at two points 3 The GSS uses a multistage probability sampling design, based on the U.S. Census. Therefore, the 1985 survey was based on the 1980 Census data, SOCIAL ISOLATION IN AMERICA 357 in time, representing the same underlying pop- ulation in 1985 and 2004. Measures We use the same measures of network charac- teristics that Marsden (1987: 123-24) used in his description of the structure of 1985 American interpersonal environments. Size is the number of names mentioned in response to the "name generator" question. Since family members and non-kin are fairly distinct institutional bases of connectedness, Marsden (1987) focused his analysis on the kin and non-kin composition of these networks. We present these comparisons, and the distribution of ties across the entire range of possible relationships measured by the GSS question (Spouse, Parent, Sibling, Child, Other family, Coworker, Member of group, Neighbor, Friend, Advisor, Other). Marsden (1987) also was concerned about the range or concentration of the interpersonal envi- ronment, recognizing the well-known fact that tightly connected, closed interpersonal envi- ronments tend to be made up of similar others and to provide fewer independent sources of information. The contrast between range and concentration also affects normative pressures — both in terms of pressure to conform and the responsibility for support in times of need. Like Marsden, we use density of the interpersonal network as an indicator of network concentra- tion (the inverse of range). It is defined as the mean intensity of tie strength among the dis- cussion partners mentioned. The GSS data are coded if the respondent reports that two of his or her confidants are total strangers, 1 if they are especially close, and 0.5 otherwise. We also include additional measures of tie strength, duration, and frequency of contact With the per- son mentioned. Tie duration was measured with a question about how long the respondent had known his or her confidant. Frequency of con- tact was measured by asking how often the Analyses We begin with an analysis that directly parallels Marsden (1987), the much-cited description of the interpersonal environments published in the American Sociological Review for the 1985 data. In each case, we first replicated Marsden's (1987) analyses on the unweighted 1985 GSS data. We then applied weights to make the data representative of the national population. 4 To describe the basic parameters of discussion net- works, we replicate the Marsden (1987) tables using appropriate weights for both 1985 and 2004. Then, we decompose the difference in core discussion network size using analyses that allow us to control for demographic changes across the two decades, to examine some pos- sible changes in reactions to the survey process, and to check for interactions of these variables with year. The negative binomial regression analysis (a change from Marsden 1987), acknowledges the fact that our dependent vari- able is a count variable. Finally, we use logistic regression analysis to distinguish social iso- lates and those who report at least one discus- sion partner. RESULTS Network Size The major finding of this study is in the first two columns of Table 1: the number of discus- 4 We note that the 1985 results in our tables differ very slightly from those of Marsden (1987). The GSS sampling frame actually selects households within blocks; the survey therefore must be weight- ed by the number of adults in the household eligible for the survey in order to constitute a representative sample of individuals in the population. Marsden (1987) presented statistics based on an unweighted sample. In 2004, the weighting scheme was slightly more complicated. After an initial round of data col- lection was completed, half of the non-contacted 358 AMERICAN SOCIOLOGICAL REVIEW Table 1 . Size of Discussion Networks, 1 985 and 2004 b Total Discussion Network Kin Network 3 Non-Kin Network 3 Network Size 1985 2004 1985 2004 1985 2004 10.0% 24.6% 29.5% 39.6% 36.1% 53.4% 1 15.0% 19.0% 29.1% 29.7% 22.4% 21.6% 2 16.2% 19.2% 21.0% 16.0% 18.1% 14.4% 3 20.3% 16.9% 11.7% 9.4% 13.2% 6.0% 4 14.8% 8.8% 5.8% 4.0% 6.8% 3.1% 5 18.2% 6.5% 2.8% 1.3% 3.4% 1.4% 6+ 5.4% 4.9% — — — — Mean 2.94 2.08 1.44 1.12 1.42 .88 Mode 3.00 .00 .00 .00 .00 .00 SD 1.95 2.05 1.41 1.38 1.57 1.40 Note: N (1985) = 1,531; N (2004) = 1,467. " Information on kinship was collected on the first five alters cited. Therefore, the sum of kin and non-kin alters is not equal to the overall network size distribution. b In all tables for this paper, cases are weighted to reflect the population. Weight variable for 1985 is a function of the number of adults in the household (ADULTS), while the weight variable for 2004 is WT2004NR. sion partners in the typical American's interper- sonal environment has decreased by nearly one person (from a mean of 2.94 to a mean of 2.08). The modal number of discussion partners has gone from three to zero, with almost half of the population (43.6 percent) now reporting that they discuss important matters with either no one or with only one other person. The decrease is espe- cially marked among those who report four or five discussion partners: these respondents have gone from a third of the population (33.0 percent) to only 15.3 percent of the population. The small number of people who report very large discus- sion networks (six or more) has decreased less markedly, from 5.4 to 4.9 percent. The next columns of Table 1 show that this marked social change has occurred in both kin and non-kin discussion partners. 5 Both have dropped from a mode of 1 to a mode of 0. Since both kin and non-kin discussion partners have gone down, the proportion kin has increased only moderately across the 19-year span. The average proportion kin has gone up from 0.49 to 0.54). Marsden's (1987) generalization that American's core discussion networks are heav- ily constituted by family still holds. All of the changes described in Table 1 are statistically significant (as is the change in pro- portion kin). The distributions on all three vari- ables differ significantly from 1985 to 2004, and the means are all significantly smaller in 2004. Indeed, it is easier to list the few things that haven 't changed: the standard deviations of all three variables have remained relatively stable, and are not different by year. Types of Relationships Table 2 looks in more detail at the types of rela- tionships that the respondents have with their confidants, to allow us to see where this large social change is occurring. The top panel shows the percentages of respondents who mentioned at least one discussion relationship of each type. Since the overall discussion network size has aone down dramatically, we exnect that the ren- SOCIAL ISOLATION IN AMERICA 359 Table 2. Respondents Who Had Various Relationships with at Least One Confidant Type of Relationship to Respondent" 1985, %(N= 1,531) 2004, %(N= 1,467) No Confidant Spouse Parent Sibling Child Other Family Member Coworker Comember of group Neighbor Friend Advisor Other Spouse is only Confidant At Least One Non-spouse Kin At Least One Non-kin Confidant 10.0 30.2 23.0 21.1 17.9 18.2 29.4 26.1 18.5 73.2 25.2 4.5 5.0 58.8 80.1 24.6** 38.1** 2 j 1 ** 14.1** 10.2** 11.8** 18.0** 11.8** 7 9** 50.6** 19 2** 31** 9.2** 42.9** 57.2** Note: The table displays, for example, "What percent of the sample mentioned a spouse/parent/etc. as a person with whom they discussed important matters?" a Since more than one type of relationship can be mentioned for any given discussion partner (e.g., a coworker can also be a co-member of a group, an advisor and a friend), the percentages do not sum to 100. ** p < .01 (two-tailed tests). only slightly (from 23.0 to 21.1 percent). Notable for their sizeable decreases are co- member of a voluntary group and neighbor, both representing types of community ties that have been stressed in the public policy debate over civic engagement (e.g., Putnam 2000). The relationships labeled "other," while small in number, are an interesting residual category. While unmarried partners are included in the spouse/partner relationship, some respondents do not consider the family of a partner to be part of the respondent's own family. So, a boyfriend's mother, a girlfriend's mother, or a partner's son- in-law appear here as an uncoded relationship type (rather than being placed by the respondent into the category "other family"). Similarly, ex- family members no longer have family status for some respondents. Respondents reported dis- cussing important matters with ex-spouses, spouse's ex-spouses, son's father, and such. Several respondents mentioned support people Since our interest in these close personal contacts is driven partly by their ability to shape flows of information, influence, and affiliation, the bottom panel of Table 2 shows the percent- ages of respondents who have networks with dif- ferent levels of reach. In addition to the large proportion of respondents who have no one to talk to, we find that the percentage of people who depend totally on a spouse for such close contact has increased from 5.0 to 9.2 percent. The proportion of people who talk to non-spouse kin (who are likely to reside outside their own household) has dropped (58.8 to 42.9 percent). The most striking drop, however, is in the per- centage of people who talk to at least one per- son who is not connected to them through kinship, a decline from 80.1 to 57.2 percent. These latter ties are the most likely to bridge socially distinct parts of the community struc- ture, since we know that marriage and family ties are more homophilous on class, religion, ial r\r nrntpcciAtiQ I cF*r\nm* -\irr\rlrf>re I t* rr ^fliP l1 0fffiKl1+D 360 AMERICAN SOCIOLOGICAL REVIEW Table 3. Structural Characteristics of Core Discussion Networks 1985 (N= l,167 a ) 2004 (N = 788 b ) Network Density <25 9.9% 7.3% .25-49 18.5% 11.8% .50-74 37.9% 39.5% >.74 33.7% 41.4% Mean .60 .66 SD .33 .33 Mean Frequency of Contact (days per year) 6-12 3.7% 3.0% > 12-52 15.3% 10.6% >52-365 81.0% 86.4% Mean 208.92 243.81 SD 117.08 114.86 Length of Association (in years) >CM1.5 12.1% 10.7% >4.5-8+ 87.9% 89.3% Mean 6.72 7.01 SD 1.34 1.00 Age Heterogeneity (standard deviation of age of alters) <5 25.8% 29.1% 5-<10 24.6% 19.7% 10-<15 24.3% 23.9% >15 25.3% 27.3% Mean 10.35 10.34 SD 6.96 8.1 Population Age Heterogeneity 20.89 18.37 Education Heterogeneity (standard deviation of alters' educations) 0-1 31.9% 34.7% >l-2.5 41.0% 45.2% >2.5 27.0% 20.1% Mean 1.77 1.48 SD 1.52 1.38 Population Educ Heterogeneity 3.59 3.17 Race Heterogeneity (Index of Qualitative Variation) 91.1% 84.5% >0 8.9% 15.4% Mean .05 .09 SD .18 .26 Population IQV .34 .53 Sex Heterogeneity (Index of Qualitative Variation) 23.8% 24.2% .01-.90 39.9% 37.6% >.90 36.3% 38.1% Mean .67 .68 SD .43 .46 t>^«..1o+;^« jn\r GO 1 no SOCIAL ISOLATION IN AMERICA 361 very densely interconnected, with mean densi- ties of 0.60 and 0.66 respectively in 1985 and 2004. This density is the average level of inter- connection among named confidants. Recalling that a code of 1 represents the confidants being closer to each other than they are to the respon- dent, these networks are quite tightly woven. This pattern was noted by Marsden (1987: 126) and remains strong in 2004. There is a slight shift toward even more interconnected networks in 2004, a pattern that is supported by analyses of frequency of contact and duration of tie. The typical respondent now sees his or her close confidant more than once a week, on average, and has known him or her for more than seven years. In general, the core discussion networks in 2004 are more closely tied to each other, are more frequently accessed, and are longer-term relationships. Even more than in 1985, the dis- cussion networks we measure in 2004 are the closest of close ties. We can also examine the character of the interpersonal environments by examining the diversity of the people mentioned as core dis- cussion partners. Table 3 also looks at the het- erogeneity of confidants in terms of age, education, race, and sex. Here, again, we see a picture of relative stability. The mean hetero- geneity of the discussion networks is signifi- cantly less than the heterogeneity of the overall population, reaffirming the well-known finding that networks are homophilous (McPherson et al. 2001). The relatively subtle changes in the diversity of the discussion networks seem to mirror the demographic changes in the popu- lation. Age and education heterogeneity have gone down in the general population, mainly because of cohort succession, and the diversi- ty of discussion networks has gone down slight- ly to reflect that fact. Racial diversity has gone up in the population (through immigration and disparate fertility rates), and has increased in dis- cussion networks as well. (Analyses not report- ed here also indicate that more people now have a rnnfiHant nf annthpr rarp That iq rp^nnnHpnt than they are to the respondent). In addition, having kin in one's network tends to increase contacts across age categories (through con- tacts with grandparents, parents, or children), educational strata (because of cohort differ- ences in educational stock), and sex (because of the heterosexual nature of marital unions and the sex composition of sibship), while it reduces heterogeneity of network ties on race, religion, geographic origin, and other matters (McPherson et al. 2001; Marsden 1987: Table 2). Comparing 1985 and 2004, we see that most of the effects of the proportion of kin in one's core discussion network on the interconnect- edness and diversity of network contacts are quite stable over the time period. Since these pat- terns are relatively well known, we present them in an Online Supplement and comment only on significant changes here (see Online Supplement on ASR Web site: http://www2. asanet.org/journals/asr/2006/toc051.html). Having kin as confidants tends to make one's network more interconnected and dense — since kin tend to know each other and perhaps be close. This effect, however, is somewhat less marked in the 2004 data than in the 1985 data. Regressing density on proportion kin produces an OLS coefficient of .26 in 2004, as compared to .36 in 1985; the proportion kin coefficient interacts significantly with year). 7 Furthermore, the predicted value of density when a network has no kin in it has increased in 2004 compared to 1985, indicating that even non-family dis- cussion partners are now more likely to know each other and be close. The effect of kin on age heterogeneity in dis- cussion networks has increased, probably because of changes in cohort structure. Networks of kin are more age diverse now than in the 1980s, while discussion networks with- out kin are more age homogeneous. The largest change, by far, is in the coefficient-related pro- portion kin in the discussion network to educa- tional hptprrmpnpit\r A/TarcHpn MQR7- Tahlp 1\ 362 AMERICAN SOCIOLOGICAL REVIEW finds a large positive OLS coefficient (A2,p < .01). The weighting by adults in the household changes this coefficient much more than most other findings, reducing the effect to .30 (still statistically significant at^ < .01). The impact of kin on educational diversity is much lower in 2004 (a coefficient of .20) and is no longer sta- tistically significant. Both kin and non-kin net- works have gotten less educationally heterogeneous by 2004 — primarily because of cohort succession and the increasing educa- tional stock of the population as a whole. The difference in cross-educational contact, poten- tially important for both the framing of issues and the flow of information, no longer varies significantly if one has only kin for confidants or no kin at all in one's discussion network. Demographic Variation in Networks Marsden (1987) also examined how important demographic categories varied in terms of their interpersonal environments. Table 4 here repro- duces some of the most important analyses shown in Marsden's (1987) Tables 3 and 4. We use OLS to see how age, education, race, and sex influence the size, kin composition, and density of one's core discussion networks. Age, which structured networks significant- ly in 1985, has very little impact on contempo- rary confidant networks. Marsden (1987: 127-28, Table 3) found a curvilinear pattern, with network size (especially non-kin confi- dants) dropping off quite precipitously with increasing age and the proportion kin being somewhat higher among younger respondents and the elderly ages. In contrast, age is not strongly related to size or kinship composition in 2004. None of the nonstandardized coeffi- cients regressing the network characteristics on age and age squared is statistically significant, and the multiple correlation between age, age squared, and network characteristics is not sig- nificantly different from zero. Clearly, there has heen cohort succession since 1 985: the verv more educated people have a lower proportion of kin in their networks than people with less education. In the confidant networks of men and women, we see that women still have significantly more kin in their networks than men do, but they no longer have fewer non-kin confidants than men. Since the size of both the kin and non-kin coef- ficients has gotten smaller from 1985 to 2004, we find that women no longer have a signifi- cantly higher proportion of kin in their net- works when compared with men. Since the kin-dominated nature of women's networks is one of the staples of the social capital literature (c.f, Moore 1990), this social change is poten- tially important. It is especially noteworthy that the shift occurs not because women are dropping kinship ties, but rather because they are achiev- ing equality with men in non-kinship ties. Unfortunately, as with growing wage equality, the equity is being achieved by men's shrinking interconnection with non-kin confidants rather than by women's greater connection to the world outside the family. Race continues to have a broad impact on net- works in American society. Both blacks and other-race respondents have smaller networks of confidants than white Americans (the reference category). This pattern is most apparent in kin- ship networks, which are markedly smaller among non-whites. A Preliminary Summary of Social Change in Networks In spite of a large literature on declining civic engagement and neighborhood involvement, we began this analysis with the expectation that networks of core confidants would be a stable feature of one's interpersonal environment. Given the close, densely interconnected nature of the ties generated by the GSS question, it seemed unlikely that the typical American would not mention several people in response. We SOCIAL ISOLATION IN AMERICA 363 & y; 1 5 a d cu z d (D i^L *: o C n z &, z z Z z z -f O CN CN r- c O m O O oc O 1 00 c 1 *C O CO CO O CN — co co ■j-. CO CO CO CO 3 3 "C Z Z z z Z Z z <U d Cfl O O c CN CN in T oc CI r- O -a O 1 O 'sC O 1 00 in O O q a V ■ n. j= 'J "3 u >-, 5 CO co CO co GO CO CO CO 3 z z z z z Z Z Z^ a c Cfl O O cc O o\ a\ O d O O 1 -t- O c 1 t- "n c O 1 O 3 OJJ CO CO GO tu d a. Cfl CO CO CO ra Z Z Z "3 > H O OV en rn t- in r- Ov OC ■ct O U O O ^D c 00 <*■ O c in O *J E u d u "5 J3 'E3 P < aj CO CO r. co CO CO CO CO CO ■A z Z Z Z Z Z Z z z 'O _S O OC O oc 00 -r CN Ov O CN — O CD O oe O CN O O — — ' O on "O 1 1 1 Eh c <L> -j q N & 73 S3 rn O r- r- m r- ts O CO Z O CO Z c -t- CO Z O 73 -P u 'cfl -3 O O — — -T -~ CO "n O ^ c -T O tj | I U Is Cl ta lX Ih ■— . ■J DO C- C- 1 O c ?3 co CO co co O "J Z Z Z Z "O Cfl O 'sC O m in CO rn rn rn d rt O O oc O t CN O in ■* CN "05 P S ■ 1 ^H 1 — "C Tj P d U Z co z O Cl N O CN O 'sO CN *Ti 00 OC 00 in CN SO U "S O q q q — q CN CN q in ■* in q ^L l" CN r f p CO 5 d 73 'Si 'O x co CO ■j-. CO tM 73 Z Z Z z Z OJ 'Eh n O In O en m in CT\ OC C ■* CN CN CJ 73 q O r q q •""S q q ~t t> q r r CN CN q d > 1 _73 D- x ■jz CO CO p TD O z CN in r^ OV CN z in CN Z O 00 Z rn rn CN 05 P 1) ■-■■ O r rn q <r* r O r [' O rn q O N 'cfl d O u S 1> c H O 2 'J •55 2 -a (L) ^^ +-» h d +-■ 00 u ed o 2 § •a ^~ p. .2 Z S 364 AMERICAN SOCIOLOGICAL REVIEW membership and neighbor have decreased dra- matically. Such a large, unexpected social change rais- es immediate questions. Therefore, in the next section we explore some reasons why the appar- ent difference between 1985 and 2004 might be artifactual. We also review other trends that might support or question our results. COULD SUCH A LARGE SOCIAL CHANGE BE REAL? Social change is best measured when bench- marks are frequent. Since our measurements are 19 years apart, we have no way to assess directly whether or not the dramatically small- er 2004 networks are part of a slowly develop- ing trend. We therefore must consider threats to validity and look at related data to see if other trends might show similar patterns. Study Design The most common threat to trend measurement is change in the questions themselves. The GSS asked the same question in 1985 and 2004. While the important matters that respondents discussed may have shifted with demographic characteristics or historical context, there is no reason to expect that the 2004 important-mat- ters question would not elicit the close, fre- quent confidants that it did in 1985. Interviewer training and probe patterns also were very sim- ilar across the two surveys. The GSS imple- mented a number of changes in sampling frame and survey procedures during the two-decade period in question, but these seem very unlike- ly to have created the observed pattern. 8 Context Question order is a vexing, important, and understudied aspect of survey design (see review in Smith 1989). Context effects are generally not large, however, and tend to be concentrated within modules of questions on similar con- tent. In methodological experiments conducted in 1988, when the GSS core questions were changed, Smith (1989) estimated that only six out of 358 questions showed real context effects. For questions like the ones of interest here, however, preceding questions can influence what one thinks of as important matters (Bailey and Marsden 1999; Bearman and Paragi 2004) and, to a lesser extent, which alters one names (Straits 2000). In 1985, the network questions were preceded by a battery of questions on reli- gion. In 2004, they were preceded by a module of questions on voluntary group membership. While not identical, the fact that a large pro- portion of the voluntary sector is composed of religiously affiliated associations (Bonikowski and McPherson 2006) means that the connec- tions that would be cognitively primed would be somewhat similar in both cases. If there were a bias introduced by this contextual feature, one suspects that it would lead to overreporting of co-membership relationships in the 2004 net- work data (since the groups of which the respon- dent and his/her alters might have been co-members had just been reviewed, and the topics that they invoked presumably primed). Recall from Table 2, however, that co-mem- bership relations declined more than other types of relations. A more serious possibility is that the volun- tary organization questions in 2004 had a train- ing effect on respondents — effectively teaching them that mentioning a larger number of affil- iations in response to an initial question would then lead to more questions about each men- tioned connection. 9 Luckily, the GSS network questions were partially repeated in 1987 in a module on political participation. In 1987, the network question appeared just after the battery of questions on voluntary association. (In this case, the network question was not followed by queries about the alter's characteristics, but instead was narrowed to a focus on political SOCIAL ISOLATION IN AMERICA 365 1985 and 2004.) In supplement data, we com- pare the limited analyses that can be replicated comparing 1987 and 2004, separated by 16 years and both preceded by voluntary associa- tion modules (see Online Supplement on ASR Web site). In these replications, we again find a dramatic drop in network size (from 2.63 in 1987 to 2.08 in 2004,/; < .01) and a dramatic increase in the proportion of respondents with no core confidants (4.5 percent in 1987 and 24.6 percent in 2004,/; < .01). There may have been some tendency for the voluntary associa- tion context effect to suppress very large net- works. Comparing the 1987 data to the 1985 data, we see fewer networks of sizes three, four, five, and more. Yet the voluntary association context decreased the number of people who reported no confidants; that proportion is actu- ally smaller in the 1987 data than in the 1985 data (4.5 percent as compared with 10.0 per- cent). The relationship between voluntary associa- tion membership and network size is positive and roughly the same size in both surveys (a cor- relation of .22 in 1987 compared to a correla- tion of .18 in 2004). This relationship is a substantively reasonable one: there is a large lit- erature on the interrelationship of networks and voluntary groups (McPherson 1983, 2004; McPherson and Ranger-Moore 1991; McPherson, Popielarz, and Drobnic 1992). Ideally, of course, one would want an experi- ment embedded in the survey design that assessed how context affected the network ques- tions. In time, such a measure of context effect should be possible. The National Science Foundation has funded a re-interview of the 2004 GSS respondents to further link their net- works and voluntary association memberships through a life history calendar (BCS 052767 1, "Niches and Networks: Studying the Co-evo- lution of Voluntary Groups and Social Networks," $746,000). These interviews will be conducted in the fall of 2006, two years after might attempt to speed the survey process along by saying that they have few (or no) entries in the list. The GSS is a long survey, lasting over an hour for many respondents. Therefore, one must be concerned with fatigue effects, espe- cially if these effects differed in 1985 and 2004. The network items occurred near the end of the survey in both years. The GSS asks a core of sociodemographic and social trend questions in each year, 10 followed by modules of questions on various topics. In 1985, the first network question was question 127 out of 148 total ques- tions. In 2004, the name generator question was also numbered 127, but this has less meaning in a CAPI survey where different questions take on different positions depending on skip pat- terns. It occurred, however, after 109 questions in the core and a module of questions about membership in voluntary associations. The GSS has the interviewer rate the coop- erativeness of the respondent immediately after the face-to-face session is completed (soon after the network questions in both years). Respondents are categorized as interested/ friendly, cooperative, restless/impatient, or hos- tile. The 2004 respondents were no more like- ly to be impatient or hostile than were the 1985 respondents (less than 4 percent in both years). The great majority of respondents were rated in the most positive category (interested/friendly) in both years (79.3 and 82.2 percent in 1985 and 2004 respectively). As we expected, cooperativeness is strongly related to the number of people who are report- ed as confidants, with hostile respondents reporting almost two fewer confidants than interested and friendly respondents. There was no statistical interaction, however, between the cooperativeness variables and survey year in predicting the number of discussion partners mentioned. To the extent that uncooperative- ness leads to underreporting of network ties, this factor seems to have operated in similar ways in both survey years. We also note that some of thp rplatinnshin hptwppn rnnnprativpnpsn and 366 AMERICAN SOCIOLOGICAL REVIEW situation of a face-to-face interview may also be more sociable in other settings. We also constructed an index of how many questions prior to the network module had miss- ing data for each respondent. Our logic was that refusal to answer preceding questions might be a behavioral indicator of fatigue or non- cooperativeness. Indeed, the number (out of 10) questions coded missing immediately prior to the network module is correlated -0.16 with the number of network alters mentioned (p < .01). We therefore control for this index of missing data in our multivariate analyses of network Convergent Data from Other Sources In the case of most major social changes, researchers can triangulate from multiple data sources at multiple time points to establish an overall pattern with some certainty. Since schol- ars have rarely measured networks in a way that can be generalized to the national population, we have fewer resources here. There are, how- ever, two types of evidence that might reinforce the data that we present. The first source of convergent data is Bearman's and Parigi's (2004) finding that 20 percent of the North Carolinians that they inter- viewed in 1997 have no one with whom they dis- cuss important matters. The proportion of people who report no confidants in the North Carolina study is consistent with the trend between the 10.0 percent estimated from the 1985 GSS sam- ple and the 24.6 percent estimated from the 2004 sample. In supplemental data, we also note that the 1987 GSS data show a movement toward a lower network size (see Online Supplement on ASR Web site). On the other hand, some telephone surveys of the national population asking questions about the number of close friends show rather different results. In 1990, for example, the Gallup Poll found that only 3 percent of their sample reported no close friends; only 16 per- cent had less than three friends. While there are many differences between the Gallup and GSS surveys, this raises the interesting question of whether the important-matters question gets at closer, core ties than the concept of close friend. Another recent telephone survey by Pew also found much larger numbers of core or close friends, when it asked about a combination of types of contact (Boase et al. 2006). Both of these surveys alert us to the possibility that respondents might be interpreting "discuss" in a literal way, and not including some types of personal contact (see Conclusions section). On the other hand, the Pew survey has a response rate of 35 percent, while the GSS consistently gets more than 70 percent of its sampled units. Our analyses (not reported here, but available from the authors) of the 2004 weights used in the GSS indicate that easily reached respon- dents are quite different from difficult-to-inter- view people in terms of their interpersonal environments. This fact reinforces the impor- tance of response rates in studies of affiliation, social networks, and civic participation. The second area where we look for conver- gence is other trend data reported by the large, hotly contested literature on civic engagement. Putnam (1995, 2000) raised the issue of declin- ing embeddedness in civic and neighborhood associations to the attention of both policy- makers and scholars (especially in political sci- ence, where networks had not been a central topic previously). While there has been sub- stantial debate about his data and the down- ward trends that they indicate (c.f, Fischer 2005; Paxton 1999; Rotolo 1999; Rotolo and Wilson 2004; Sampson 2004), the decline that he reports in socializing among neighbors and general participation in social life beyond the level of the nuclear family fits well with our observations that association co-members, 11 Since the index considers different questions in n piohhnr« and pvtpndpd fqmilv arp mpntinnpd SOCIAL ISOLATION IN AMERICA 367 to the social changes that we observe. For exam- ple, the decline in socializing with neighbors has been about 3 percent over the past two decades. Respondents in 2004 are somewhat less likely than those in 1985 to report that they can trust other people, think that they are fair (as opposed to taking advantage), and think that they are helpful (as opposed to looking out for them- selves). The changes in these variables, howev- er, are in the order of 2 percent (fair) to 9.6 percent (helpful) — again, small relative to the drop that we see in core network size. Demographic Change as a Source of Network Change Of course, the demographic characteristics of the country have changed considerably in the two decades. Some of those changes could have resulted in a shrinking network size even in the absence of non-demographic social change. As the population gets older and more racially diverse, we would expect networks to get small- er, since older people and racial minorities have smaller networks, on average. On the other hand, the increasing education of the population should tend to increase network size. To assess the extent to which basic demographic changes have altered the landscape of interpersonal envi- ronments, we now move to a multivariate model to examine change from 1985 to 2004. CHANGE NET OF DEMOGRAPHIC AND METHODOLOGICAL FACTORS We use negative binomial regression to model the size of discussion networks, because our dependent variable is a count of network alters. Here, data from both the 1985 and 2004 GSS are combined, with the survey year acting as an independent variable in the analysis. Table 5, Model I, illustrates the most important social change documented by our earlier analyses of discussion networks: the number of confidants has decreased significantly over the period between the two surveys. This negative binomial coefficient of -.356 (evaluated with the Y-inter- cept) corresponds to a drop of .86 network alters by 2004 (c.f, row 8 of Table 1, results round- ed). The coefficients in all models for Wave Table 5. Multivariate Models of Discussion Network Size and Social Isolation Model Dependent Variable Dependent Variable: Discussion Network Size No Discussion Partners (Negative Binomial Regression) (Logistic Regression) Independent Variable I II III IV V Constant 1.078 1.150 .477 -2.144 -1.297 Wave(l =2004) -.356 -.329 -.407 1.374 214NS, b Cooperative (Compared to Friendly/Interested) — -.225 -.145 .126 NS .132 NS Restless/Impatient — -.667 -.585 1.295 1.308 Hostile — -1.121 -.985° 2.005 2.016 Number Missing in Previous Module — -.257 -.198 .372 .376 Education (in yrs) — — .059 -.087 -.158 Education* Wave — — — — .099 Female — — .071 c -.194 NS -.195 NS Age a — — -.002 .016 .015 Currently Married — — .061 c -.256 -.253 368 AMERICAN SOCIOLOGICAL REVIEW (the 1985-2004 contrast) are a test of the null hypothesis that differences in network size between the two surveys are due to sampling error. Model II adds the indicators of fatigue and hostility that we suspect may lead respondents to underreport their network ties. The more hos- tile the respondent gets, the more he or she is likely to report a small network. Having miss- ing data on questions that precede the network questions serves as an additional indicator of survey problems. Controlling for these data issues does not, however, significantly reduce the drop in discussion network size from 1985 to 2004. Controlling for demographic factors actual- ly increases the estimated difference in net- work size over the 19-year period (Model III). This effect occurs because education is posi- tively associated with network size, and educa- tional levels have increased over time. This effect more than offsets the declines in network size due to other factors such as the declining proportion of the population that is married and the growing minority population. More educated and younger people have sig- nificantly larger discussion networks, as do women. Network size gradually shrinks with aging, and non-white Americans have fewer network resources. Marriage draws one into networks of people with whom one discusses important matters (notably one's spouse, the most often-named type of relationship for the discussion partner). 12 Of course, there are many controls that we could implement. The results in Table 5 repre- sent the major, stable, statistically significant demographic sources of confidant networks. Some of the logically plausible socio-demo- graphic variables are not important sources of network variation in these data. For example, the opportunity structures represented by number of siblings, number of children, and number of adults in the household do not significantly affect the number of confidants (possibly resented as hours worked per week or as dummy variables for full-time and part-time work) does not have an effect. Geographic mobility does not appear to have an impact, although our ability to explore this factor is limited by the fact that the "size of place" variable has not yet been added to the 2004 GSS. 13 Neither size of place of residence at age 16 nor whether or not the respondent has moved geographically since age 16 has an effect. While a full exploration of the non-demographic sources of confidant networks is beyond the scope of this article, some com- monly used predictors like the number of hours spent watching television are also unimportant (Putnam 2000). Therefore, we conclude that the large drop in confidant networks between 1985 and 2004 in these data is unlikely to be a result of population shifts on other variables. Since negative binomial regression coeffi- cients are not as intuitively interpretable as OLS coefficients, we offer the following predicted values from Table 5 as illustration of our main result. In 1985, a white married 25-year-old male high school graduate who was an inter- ested, friendly respondent to the survey, and who had no missing data on any of the 10 items preceding the network module, would be expect- ed to have slightly more than three confidants (3.3) with whom he discussed important mat- ters. In 2004, an interested, friendly fellow with the same demographic characteristics would have reported a network more than one alter smaller (2.2). Another way of viewing the same comparison would be to age our friendly fellow by the 19 years of the study (from 25 to 44), leading to an even smaller network of 2. 1 alters. The resources represented by core networks mirror other major class divides in our society. Net of all other factors, increasing education sharply increases the number of discussion part- ners that a respondent reports, from roughly 1.5 alters for a person with the lowest level of education in 1985, to around five alters for such a person at the highest level of education. The differences for 2004 are smaller, but iust as SOCIAL ISOLATION IN AMERICA 369 alters. The differences between 1985 and 2004, however, remain salient even in the face of this major divide. In 1985, high school dropouts (with 10 years of education) had a network with roughly 2.8 discussion partners — in the range of a college graduate in 2004. Figure 1 brings these stark differences in educational trajectory to bear on the issue of kin composition and network range. In both time periods, education promotes discussion with both kin and non-family members, with ties outside the family affected most markedly. The different slopes of these two curves mean that at some level of educational achievement the two curves cross. Discussion networks become dominated by people outside one's immediate family. In 1985, this cross occurred at around 13 years of education, a little more than a high school diploma. Discussion networks of those with some college comprised more non-family than family; college graduates had more confi- dants outside their kin group than inside it. In 2004, the non-kin ties have dropped so much that this crossover is not predicted to occur until a respondent has acquired post-college educa- tion. Clearly, the net effect of these changes is to focus and limit the reach of core discussion networks in the general population. THE SHAPE OF SOCIAL ISOLATION Given the close, dense nature of core discussion networks, one might argue that the crucial dis- tinction is not among different network sizes but between those who have someone to talk to and those who report no one with whom they can discuss matters that are important to them. In Table 5, Model IV, we present a logistic regres- sion analysis that contrasts those who did not name anyone in answer to the name generator (even after the obligatory probe by interview- ers) and those who did name a discussion part- ner. Most of the effects are what we would have expected from our earlier analysis of network size. The data issues operate in the same man- ner — more cooperative respondents are less likely to be socially isolated, while those who having lots of missing data are more likely to be isolates. More highly educated, younger, cur- rently married people are less likely to be social isolates. The only notable change from Model III here is that men are not significantly more 3.5 2.5 1 1.5 0.5 370 AMERICAN SOCIOLOGICAL REVIEW likely than women to be social isolates in core discussion networks. They may have fewer dis- cussion partners than women, but they nonetheless are as likely to have at least one confidant. Similarly, other-race people are not significantly different from whites (although blacks are still more likely than whites to be isolates). In the analyses reported in Table 5, we test- ed for all possible two-way interactions between survey year and predictor variable. In the logis- tic regression analysis of the probability of being a social isolate, we found an interaction between survey year and education. Model V shows that interaction. The effect of education on the probability of being a social isolate is strong and negative in 1985 (a coefficient of - 0. 158), and becomes somewhat less negative in 2004 (-0.158 + 0.099 = -0.059). Again, to make the logistic regression coef- ficients somewhat more vivid, we compute the predicted probability that our white married 25-year-old male high school graduate who is enthusiastically participating in the survey and leaving no missing data would have someone to talk to about important matters in 1985. He would be virtually assured of a discussion part- ner (predicted probability of being an isolate = 0.04). The same type of person in 2004 would have a more a ten percent chance of being an isolate (predicted probability = 0.16). Do the same mental experiment and age our 25-year- old to 44 years of age in 2004: we find that his probability of being an isolate would have quin- tupled from 0.04 to 0.20. UNEVENNESS IN THE SOCIAL CHANGE Given that social change rarely affects the entire population simultaneously, the relative lack of interactions seems somewhat strange. We there- fore explore in more depth possible uneven- ness in the network changes that we observe. While the change is unusually pervasive (prob- ably because of the 19-year gap in our assess- ment), there are some hints about uneven change in different social groups. First, there is the statistical interaction between education and year in affecting the probability of social isolation. This interaction is made clearer by inspection of Figure 2, which 0.35 0.3 3 0.25 1 0.2 o 0.15 0.1 0.05 SOCIAL ISOLATION IN AMERICA 371 plots the fitted probability of social isolation across years of education in 1985 and 2004. As the figure shows, there is a very sharp increase in the probability of social isolation for all lev- els of education, but the greatest change occurs in the middle range of education. In 1985, increasing education led to a sharp decline in social isolation, while that effect is much less evident in 2004. This change is one of the few areas where inequality has gone down in our society. Unfortunately, the inequality is decreas- ing because everyone is getting worse off (if we assume that social isolation is bad). We also inspected the data for interesting subgroup differences, using the intersection of race, class, and gender as a general guide. The decline in networks is quite uniform, but young (ages 18-39), white, educated (high school degree or more) men seem to have lost more dis- cussion partners than other population groups (from 3.5 in 1985 to 2.0 in 2004). In the next section, we discuss the possible impact of Internet usage on this demographic group. Young, poorly educated (less than high school), white women also experienced a large decline (3.2 to 1.4 alters). Among African Americans, a gender differ- ence is striking. Older (60+) African American men's networks have declined the most (from 3.6 to 1 .8). Among black women, the change is more uniform, with the young experiencing a larger decline than the old. Indeed, black men over 60 are the only sector of the older popula- tion that experienced a major decline between 1985 and 2004. Otherwise, the elderly have been more stable than most other groups in their core social connections. 14 This article leaves these possible subgroup changes in core discussion networks to future analyses. DISCUSSION AND CONCLUSIONS If we assume that interpersonal environments are important (and most sociologists do), there appears to have been a large social change in the past two decades. The number of people who have someone to talk to about matters that are important to them has declined dramatically, and the number of alternative discussion part- ners has shrunk. In his groundbreaking study of social networks, To Dwell among Friends, Claude Fischer (1982:125-27) labeled those who had only one or no discussion ties with whom to discuss personal matters as having marginal or inadequate counseling support. By those criteria, we have gone from a quarter of the American population being isolated from counseling support to almost half of the popu- lation falling into that category. The American population has lost discussion partners from both kin and outside the family. The largest losses, however, have come from the ties that bind us to community and neighbor- hood. The general image is one of an already densely connected, close, homogeneous set of ties slowly closing in on itself, becoming small- er, more tightly interconnected, more focused on the very strong bonds of the nuclear family (spouses, partners, and parents). The education level at which one is more connected through core discussion ties to the larger community than to family members has shifted up into the graduate degrees, a level of education attained by only a tiny minority of the population. High school graduates and those with some college are now in a very family-dominated social envi- ronment of core confidants. Some of the basic parameters of discussion network structure have moved very little in 1 9 years. Age and sex heterogeneity of ties has remained remarkably constant, and the decline in educational diversity seems directly linked to the increasing education level of the population. Racial contact in these discussion networks has actuallv increased. Havine a network dominat- 372 AMERICAN SOCIOLOGICAL REVIEW now kin and non-kin look similar in their edu- cational composition. If core discussion networks represent an important social resource, Americans are still stratified on education and race. Higher edu- cation people have larger networks of both fam- ily and non- family members, and their networks have more of the range that tends to bring new information and perspective into the interper- sonal environment. Non-whites still have small- er networks than whites. Sex, on the other hand, seems to have lost some of its interpersonal stratifying power in the past 19 years. While women still have marginally larger networks than men and have more discussions about important matters with kin, they no longer show a significant deficit in the number of core con- tacts outside the family. As a result, women no longer have a significantly more kinship- focused discussion network than men; nor are they significantly less likely than men to be social isolates. Our final estimates, corrected for response problems and demographic shifts, are that (1) the typical American discussion network has slightly less than one fewer confidant in it than it did in 1985, and (2) that in 2004 an adult, non- institutionalized American is much more like- ly to be completely isolated from people with whom he or she could discuss important mat- ters than in 1985. Given the size of this social change, we remain cautious (perhaps even skep- tical) of its size. The limited network data in 1987 indicate that the proportion of people who answer "no one" and who list relatively large numbers of confidants may be especially sen- sitive to context effects (see Online Supplement on ASR Web site). Given our analyses of the highest-quality nationally representative data available, however, our best current estimate is that the social environment of core confidants surrounding the typical American has become smaller, more densely interconnected, and more centered on the close ties of spouse/partner. Thp tvnp« nf hriHaina tipe that mnnppt iiq tn other literature) to guide future research. Three explanations seem most likely. The first two possibilities concern how peo- ple interpret the question that we asked them, in view of historical and cultural change. What Americans considered important might well have shifted over the past two decades, perhaps as a result of major events (the attacks of 9/1 1 and the wars that followed). If people think of "important" more in terms of national and world-level events, more people might now think that they have nothing important to say. 15 Since many people interpret the question as simply asking about their close confidants (rather than a particular discussion of important matters), it seems unlikely that such a shift in cultural meaning would have produced such a strong effect. It may, however, have contributed to the pattern. The second possibility is that the use of the word "discuss" in the question was interpreted by respondents to exclude other forms of com- munication that are becoming dominant in our contacts with core confidants. Many more peo- ple now use cell phones and Internet (email, list serves, chat rooms, and instant messaging) to contact core network members (Wellman et al. 2006; Boase et al. 2006). If people exclude these types of communications when answering the question, it could reduce the number of alters reported. 16 The third possibility is the most substantive- ly interesting. Shifts in work, geographic, and recreational patterns may have combined to cre- ate a larger demarcation between a smaller core of very close confidant ties and a much larger array of less interconnected, more geographi- 15 Bearman and Parigi (2004) found that roughly half of their respondents who reported discussing important matters with no one in the past six months said that they had nothing to say. 16 The fact that cell phones and Internet commu- nications tend to mirror other channels of commu- SOCIAL ISOLATION IN AMERICA 373 cally dispersed, more unidimensional relation- ships. Families, especially families with chil- dren, may face a time bind that comes from longer commutes and more work time (Hochschild 1997). As more women have entered the labor force, families have added 10 to 29 hours per week to their hours working out- side the home (Jacobs and Gerson 2001; Hout and Hanley 2002). The increase has been the most dramatic among middle-aged, better-edu- cated, higher-income families — exactly the demographic group that fuels the voluntary association system (McPherson 1983; McPherson and Ranger-Moore 1991). The nar- rowing of the education gap suggests that this group — highly educated middle-class fami- lies — is where the declines in the number of core discussion ties have been sharpest. Such fami- lies can use new technologies to stay in touch with kin and friends — most notably cell phones and the Internet. While these technologies allow a network to spread out across geographic space and might even enhance contacts outside the home (e.g., arranging a meeting at a restaurant or bar), they seem, however, to lower the prob- ability of having face-to-face visits with fami- ly, neighbors, or friends in one's home (Boase et al. 2006; Gershuny 2003; 2and Erbring 2000; Nie, Hillygus, and Erbring 2002). 17 Wellman et al. (2006:10-13) note that Internet usage may even interfere with communication in the home, creating a post-familial family where family members spend time interacting with multiple computers in the home, rather than with each other. They suggest that computer technology may foster a wider, less-localized array of weak ties, rather than the strong, tightly intercon- nected confidant ties that we have measured here. This may not be all bad, of course, since we know that weak ties expose us to a wider range of information than strong, close ties. We also know, however, that strong ties offer a wider array of support, both in normal times (Wellman and Worlev 1990*1 and in emergencies fHurlhurt et al. 2000). Only geographically local ties can offer some services and emotional support with ease (Wellman and Worley 1990). Whatever the reason, it appears that Americans are connected far less tightly now than they were 19 years ago. Furthermore, ties with local neighborhoods and groups have suf- fered at a higher rate than others. Possibly, we will discover that it is not so much a matter of increasing isolation but a shift in the form and type of connection. Just as Sampson et al. (2005) discovered a shift in the type of civic partici- pation, and the Pew Internet and American Society Report (Boase et al. 2006) showed a shift in modes of communication, the evidence that we present here may be an indicator of a shift in structures of affiliation. Miller McPherson is Professor of Sociology at the University of Arizona and Research Professor of Sociology at Duke University. His current projects include a test of his evolutionary model of affiliation with nationally representative data funded by the Human and Social Dynamics Initiative at the National Science Foundation. The project will create a representative sample of voluntary groups and study the co-evolution of group memberships and networks over time. Lynn Smith-Lovin is Robert L. Wilson Professor of Sociology at Duke University. She received the 2006 Cooley-Mead Award from the ASA s Social Psychology Section and the 2005 Lifetime Achievement Award in the Sociology of Emotions. Her research examines the relationships among social association, identity, action, and emotion. Her cur- rent projects involve an experimental study of justice, identity, and emotion as well as work with McPherson on an ecological theoiy of identity (both funded by the National Science Foundation). Matthew E. Brashears is a Ph.D. candidate in Sociology at the University of Arizona. A past win- ner of the Pacific Sociological Association and the Social Psychology Section s Graduate Student Paper Awards, he is interested in social networks and their role in information transmission and transformation. His dissertation focuses on examining the reciprocal effects of attitudinal similarity and network formation. 374 AMERICAN SOCIOLOGICAL REVIEW and Interview Context: Examining the General Social Survey Name Generator Using Cognitive Methods." Social Networks 2 1 :287-309 . Bearman, Peter S. and James Moody. 2004. "Adolescent Suicidally." American Journal of Public Health 94:89-95. Bearman, Peter and Paolo Parigi. 2004. "Cloning Headless Frogs and Other Important Matters: Conversation Topic and Network Structure." Social Forces 83:535-57. Bernard, H. R., P. D. Killworth, and L. Sailer. 1982. "Informant Accuracy in Social Network Data: An Experimental Attempt to Predict Actual Communication from Recall Data." Social Science Research 11:30-66. Boase, Jeffery, John Horrigan, Barry Wellman, and Lee Rainie. 2006. "The Strength of Internet Ties." Washington, DC : Pew Internet and American Life Project, January. Bonikowski, Bart and Miller McPherson. 2006. "Voluntary Associations." In Handbook of 21st Century Sociology, edited by Dennis L. Peck and Clifton Bryant. Thousand Oaks, CA: Sage. Brewer, D. D. 1995. "The Social Structural Basis of the Organization of Persons in Memory." Human Nature 6:379^103. Burt, Ronald S. 1984. "Network Items and the General Social Survey." Social Networks 6:293-339. . 1986. "A Note on Socio-metric Order in the General Social Survey Data." Social Networks 8:149-74. 1997. "A Note on Social Capital and Network Content." Social Networks 19:355-73. Fischer, Claude S. 1982. To Dwell among Friends: Personal Networks in Town and City. Chicago, IL: University of Chicago Press. . 2005 . "Bowling Alone: What's the score?" Social Networks 27: 155-67. Freeman, L. C, A. K. Romney, and S. C. Freeman. 1987. "Cognitive Structure and Informant Accuracy." American Anthropologist 89:311-25. Gershuny, Jonathan. 2003. "Web Use and Net Nerds: A Neofunctionalist Analysis of the Impact of Information Technology in the Home." Social Forces 82:141-68. Granovetter, Mark. 1973. "The Strength of Weak Ties." American Journal of Sociology 78: 1360-80. Hill, R. A. and R. I. M. Dunbar. 2003. "Social Network Size in Humans." Human Nature Beggs. 2000. "Core Networks and Tie Activation: What Kinds of Routine Networks Allocate Resources in Nonroutine Situations?" American Sociological Review 65:598-618. Jacobs, Jerry A. and Kathleen Gerson. 2001. "Overworked Individuals or Overworked Families." Work and Occupations 28:40-63. Kogovsek, Tina and Anuska Ferligoj. 2004. "The Quality of Measurement of Personal Support Subnetworks." Qualify and Quantity 38:517-32. Lin, Nan. 2001. Social Capitol: A Theory of Social Structure and Action. New York: Cambridge University Press. Marin, A. 2004. "Are Respondents More Likely to List Alters with Certain Characteristics? Implications for Name Generator Data." Social Networks 26:289-307 . Marsden, Peter V 1987. "Core Discussion Networks of Americans." American Sociological Review 52:122-31. . 2005. "Recent Developments in Network Measurement." Pp. 8-30 in Models and Methods in Social Network Analysis, edited by Peter J. Carrington, John Scott, and Stanley Wasserman. New York: Cambridge University Press. McPherson, Miller. 1983. "An Ecology of Affiliation." American Sociological Review 48:519-32. . 2004. "A Blau Space Primer: Prolegomenon to an Ecology of Affiliation." Industrial and Corporate Change 13: 263-80. McPherson, Miller, Pamela A. Popielarz, and Sonja Drobnic. 1992. "Social Networks and Organizational Dynamics." American Sociological Review 57:153-70. McPherson, Miller and James Ranger-Moore. 1991. "Evolution on a Dancing Landscape: Organizations and Networks in Dynamic Blau Space." Social Forces 10:19^X2. McPherson, Miller and Lynn Smith-Lovin. 1981. "Women and Weak Ties: Sex Differences in the Size of Voluntary Associations." American Journal of Sociology 87:883-904. McPherson, Miller, Lynn Smith-Lovin, and James M. Cook. 2001. "Birds of a Feather: Homophily in Social Networks." Annual Review of Sociologv 27:415^14. Moore, Gwen. 1990. "Structural Determinants of Men's and Women's Personal Networks." American Sociological Review 55:726—35. SOCIAL ISOLATION IN AMERICA 375 Paxton, Pamela. 1999. "Is Social Capital Declining in the United States? A Multiple Indicator Assessment." American Journal of Sociology 105:88-127. Putnam, Robert D. 1 995 . "Bowling Alone : America's Declining Social Capital." Journal of Democracy 6:65-78. . 2000. Bowling Alone: The Collapse and Revival of American Community. New York: Simon and Schuster. Rotolo, Thomas. 1999. "Trends in Voluntary Association Participation." Nonprofit and Voluntary Sector Quarterly 28:199-212. Rotolo, Thomas and John Wilson. 2004. "What Happened to the 'Long Civic Generation'? Explaining Cohort Differences in Volunteerism." Social Forces 82 : 1 09 1 - 1 2 1 . Ruan, Danching. 1998. "The Content of the General Social Survey Discussion Networks: An Exploration of the General Social Survey Discussion Name Generator in a Chinese Context." Social Networks 20:247-64. Sampson, Robert J. 2004. "Neighborhood and Community: Collective Efficacy and Community Safety." New Economy 1 1 : 106-13. Sampson, Robert J. and John H. Laub. 1990. "Crime and Deviance over the Life Course: The Salience of Adult Social Bonds." American Sociological Review 55:609-27 . Sampson, Robert J., Doug McAdam, Heather Maclndoe, and Simon Weffer-Elizondo. 2005. "Civil Society Reconsidered: The Durable Nature and Community Structure of Collective Civic Action." American Journal of Sociology. 111:673-714. Smith, Tom W 1989. "Thoughts on the Nature of Context Effects." Methodological Report No. 66, National Opinion Research Center, September, Chicago, IL. Smith-Lovin, Lynn and Miller McPherson. 1993. "You Are Who You Know: A Network Perspective on Gender." Pp. 223-51 in Theory on Gender/Feminism on Theory, edited by Paula England. New York: Aldine. Straits, Bruce C. 2000. "Ego's Important Discussants or Significant People: An Experiment in Varying the Wording of Personal Network Name Generators." Social Networks 22: 123^40. Verbrugge, L. M. 1977. "The Structure of Adult Friendship Choices." Social Forces 56:576-97. Wellman, Barry. 1979. "The Community Question: The Intimate Networks of East Yorkers." American Journal of Sociology 84: 120 1-3 1 . Wellman, Barry, Bernie Hogan, Kristen Berg, Jeffery Boase, Juan-Antonio Carrasco, Rochelle Cote, Jennifer Kayahara, Tracy L.M. Kennedy and Phuoc Tran. 2006. "Connected Lives: The Project." In Networked Neighborhoods, edited by Patrick Purcell. London: Springer. Wellman, Barry and Scot Wortley. 1990. "Different Strokes from Different Folks: Community Ties and Social Support." American Journal of Sociology 96:558-88.